The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits deceptive documentation practices by including numerous reference files (e.g., references/architecture-patterns.md, references/security-best-practices.md, references/state-management.md) that consist of 150+ redundant sections of repetitive boilerplate. This context flooding can be used to saturate the agent's context window, potentially hiding important information or disrupting the agent's ability to follow other instructions.
[PROMPT_INJECTION]: The skill uses authoritative framing in its documentation (e.g., 'strictly intended for Staff+ Engineers') and repeats specific technical instructions thousands of times across the file set. This technique can be used to bias the model's output or influence its behavior through repetitive conditioning.
[PROMPT_INJECTION]: The skill defines pipelines that ingest untrusted visual data (images and videos) through preprocessing and detection functions (e.g., preprocess_image in references/cv-pipeline.md). Without boundary markers or sanitization, this creates a surface for indirect prompt injection where instructions embedded in visual media could influence agent actions.
[DATA_EXFILTRATION]: Numerous files contain code snippets for concurrent network requests using the aiohttp library. This represents a functional capability for network access that could be misused for data exfiltration if the agent is manipulated into transmitting sensitive data.
[EXTERNAL_DOWNLOADS]: The skill provides instructions and examples for downloading and executing pretrained machine learning models and weights from external sources such as Ultralytics (YOLOv8), Hugging Face (DETR), and Meta AI (SAM). While these are common services, they involve the runtime execution of external binary artifacts.

ml-computer-vision